By: someone (someone.delete@this.somewhere.com), July 25, 2012 10:06 am
Room: Moderated Discussions
aaron spink (aaronspink.delete@this.notearthlink.net) on July 25, 2012 9:08 am wrote:
> sf (pc.delete@this.email.com) on July 25, 2012 6:46 am wrote:
>
> > For GPU
> compute, a CPU is needed to be "driving".
> >
> > What will
> > the
> graph look like if the driver is included in the calculation?
> >
> > What
> is the
> > currently best ratio of GPU to CPU?
> >
> > Will the best
> future design be a
> > BlueGene/Q (for best I/O) driving an optimum number of
> GPUs (for best Compute)?
> >
> Well one thing not taken into account that
> people need to be away of is effective computational efficiency. For instance,
> CPU based supers tend to be in the 85-90% range of peak while GPU based supers
> tend to be in the 50% range.
That's best case performance in dense algebra kernels in hand written
math libraries etc.
On real applications (weather simulation, circuit simulation etc) a general
purpose MPU is doing well to hit 10% of peak FLOPs. Presuming anyone
would expend the monumental effort to entirely port such apps to a GPU I
would be shocked if it could reach 1%.
> sf (pc.delete@this.email.com) on July 25, 2012 6:46 am wrote:
>
> > For GPU
> compute, a CPU is needed to be "driving".
> >
> > What will
> > the
> graph look like if the driver is included in the calculation?
> >
> > What
> is the
> > currently best ratio of GPU to CPU?
> >
> > Will the best
> future design be a
> > BlueGene/Q (for best I/O) driving an optimum number of
> GPUs (for best Compute)?
> >
> Well one thing not taken into account that
> people need to be away of is effective computational efficiency. For instance,
> CPU based supers tend to be in the 85-90% range of peak while GPU based supers
> tend to be in the 50% range.
That's best case performance in dense algebra kernels in hand written
math libraries etc.
On real applications (weather simulation, circuit simulation etc) a general
purpose MPU is doing well to hit 10% of peak FLOPs. Presuming anyone
would expend the monumental effort to entirely port such apps to a GPU I
would be shocked if it could reach 1%.
